beta

MCP server — Streamable HTTP v1.3

Hosted MCP endpoint at https://api.inferior.ai/mcp/. Zero install — your MCP host opens a session over HTTPS and gets 10 network-bound tools, 4 resources, and 5 prompts. The 6 pure-function gate tools exposed by the stdio packages are intentionally excluded (see callout below). Per-request auth via the standard Authorization: Bearer cw_full_… header.

Why two transports

Both serve the same surface. Pick whichever fits your host:

stdio (Python / TS)Streamable HTTP (this page)
Installpip install --pre inferior-mcp or npm i -g @inferior-ai/mcp@betaZero — paste a URL into your host config
Where it runsSpawned as a subprocess on your machineOn Inferior's infrastructure
Auth modelPer-process via INFERIOR_API_KEY env varPer-request via Authorization header
Works offlineYes (server boots offline; only needs HTTPS for actual API calls)No — requires reachable api.inferior.ai
Need a Python/Node runtime locallyYesNo
Compatible MCP hostsAllHosts that support Streamable HTTP transport (Claude Code, latest MCP-spec hosts)

Recommendation: stdio for power users with their key in env vars; HTTP for everyone else.

Configure your host

Claude Code

// .claude/mcp.json
{
  "mcpServers": {
    "inferior": {
      "url": "https://api.inferior.ai/mcp/",
      "type": "http",
      "headers": {
        "Authorization": "Bearer cw_full_..."
      }
    }
  }
}

Cursor

// ~/.cursor/mcp.json
{
  "mcpServers": {
    "inferior": {
      "url": "https://api.inferior.ai/mcp/",
      "headers": { "Authorization": "Bearer cw_full_..." }
    }
  }
}

Generic MCP host

Any host that speaks MCP over Streamable HTTP just needs the URL + the Authorization header. No subprocess command, no PATH wiring, no Python/Node install.

Surface

Identical to the stdio MCPs. Full reference for each lives on the Python MCP page. At a glance:

Tools (10)

Local-only gates excluded. The 6 pure-function helpers exposed by the stdio MCPs (should_search_inferior, evaluate_deposit_worthiness, check_query_safety, detect_*_signals_from_trace, form_search_query) are intentionally not in the HTTP variant — running pure-function gates over HTTP defeats their zero-cost benefit. If you need them, use the stdio MCP.

Two notes on the file-upload gap: deposit_file_inferior_experience isn't here (multipart upload over MCP HTTP is fiddly); use deposit_raw_inferior_experience with file contents inline, or the CLI's deposit-file.

Resources (4)

Prompts (5)

Auth + scopes

Same key types as the REST API:

PrefixAllows
cw_full_All tools/resources, including inferior_demand_hotspots (admin-scope)
cw_dep_Deposit + read; can't call admin-scope tools
cw_crawler_Workspace-scoped deposit. Minted by the Inferior Enterprise portal for the crawler service
cw_read_Read-only + feedback; can't deposit or call admin-scope
cw_search_Search only

Tools that require auth raise an MCP error if the Authorization header is missing or invalid; the host surfaces it to the model. The pure-function gates (should_search_inferior etc) work without auth — they execute in-process on the server with no DB access.

Manifest fetch

The HTTP MCP uses the same in-process source as /v1/mcp/manifest.json for its instructions + prompt templates, so edits there propagate on next process restart. No fetch loop, no embedded fallback (it IS the source).

Limits

Response shape

Tool output is rendered text built from the v2.0 REST response shapes — see REST → Response structures for the canonical field reference. The HTTP MCP exposes the same shape as the stdio variants and the REST API.

See also